When AI gets cheap enough, does the hiring question stop being 'who is best?' and become 'what does this cost per task?'
Google just halved the price of its coding model and OpenAI answered with a speed tier the same day — inference is now the cheapest it has been all year. If the decisive number shifts from how smart a model is to what a finished task costs, which kinds of work get handed to a machine first, and does anyone actually pocket the savings?
Commentaires (1)
In the August 14, 2026 episode of Minds, Bodies, and Terawatts, we dug into a week where the labs stopped selling genius and started selling invoices — Gemini 3.7 Flash at seventy-five cents per million input tokens, three weeks after the model it replaced. The uncomfortable part isn’t that a one-person shop now gets frontier-grade coding help for pocket change; it’s that the same price sheet sits on the desk of whoever decides whether to hire that shop at all. And the savings are less obvious than they look: agentic models burn five to thirty times more tokens per task, so a falling unit price can still mean a rising bill. Where the line lands depends less on capability than on what a task costs end to end — which is exactly the substitution threshold the book describes. Give the episode a listen, then tell us where you think the line falls in your own field.
Related reading on unscarcity.ai:
Envie d'aller plus loin ?
Obtenez le plan complet dans <em>L'ère de la post-pénurie : Repenser la société à l'ère des machines</em>
